RocmReader: GPU OCR + Cited Document RAG Two containerised AI pipelines that run entirely on an AMD Instinct/Radeon GPU with ROCm and PyTorch, using open-source Qwen3-VL vision-language models, with no external APIs or network access at inference time.
Mini-Challenge 2 (OCR): a resident GPU worker loads Qwen3-VL once and serves every image over a Unix socket, so each call is a fast round-trip rather than a model reload. The prompt follows the transcription conventions: plate numbers without state banners or slogans, province characters kept on Chinese plates, and multi-line road signs read top to bottom. PNG, JPEG and 16-bit TIFF are handled, with resizing and a contrast-boosted retry for dark, glary images.
Mini-Challenge 3 (RAG): a fault-tolerant indexer parses PDF, DOCX tables, multi-sheet XLSX, CSV, logs, source code and images (OCR'd by the same VLM), skipping encrypted, unreadable and unknown files without stopping. Questions are answered with BM25 retrieval plus identifier-aware expansion and a second retrieval hop for chained evidence (log, then ticket, then fix version). A citation verifier keeps only the files strictly necessary, drops withdrawn revisions, and refuses rather than guesses when the answer is not actually in a readable document.
Match 4 (Intelligent Web Information Retrieval): a best-first same-site crawler plus the same cited-answer engine returns the exact value and its source URL. Match 4 Image: docker.io/vrishintej/web-7c41d2a9:v1
.png&w=640&q=75)